pocket-tts is a lightweight text-to-speech application with voice cloning that runs efficiently on CPUs. It generates audio locally without requiring GPUs or external web APIs.
Project overview
By running a 100M-parameter model on just two CPU cores with roughly 200 ms latency, pocket-tts removes the GPU and external API dependencies that typically complicate local speech generation.
Project type
Audio & Speech
Use cases
Audio & Speech
Deployment
Refer to project documentation
License
MIT
Best for
Developers needing a local text-to-speech and voice cloning solution that runs efficiently without GPU hardware or paid external APIs.
Key capabilities
Generates audio speech from text inputs efficiently on standard CPUs.
Supports cloning voices from plain wav file inputs or pre-exported safetensors files.
Supports english, french, german, portuguese, italian, and spanish languages.
Provides low latency (~200ms) streaming of generated audio.
Provides a serve command to run a local server with a web interface for generating audio via HTTP requests.
Can handle extremely long text inputs for audio generation.
Allows usage as a simple Python library to generate audio from text.
Limitations and risks
Running this TTS model on the GPU did not observe a speedup compared to CPU execution.
Does not currently support adding silence in the text input to generate pauses.
Voice cloning introduces the risk of impersonation or cloning without explicit and lawful consent. Users must ensure they have proper authorization before cloning any voice.
Getting started
Install via pip install pocket-tts or uvx pocket-tts, then run pocket-tts generate to output a wav file.
Evidence and sources
README: A lightweight text-to-speech (TTS) application designed to run efficiently on CPUs. Forget about the hassle of using GPUs and web APIs serving TTS models. With Kyutai's Pocket TTS…
README: You can also use `pip install pocket-tts` to install it manually.
README: * Runs on CPU * Small model size, 100M parameters * Audio streaming * Low latency, ~200ms to get the first audio chunk * Faster than real-time, ~6x real-time on a CPU of MacBook A…
Release: v2.1.0
README: The `--voice` argument can also take a plain wav file as input for voice cloning.